Sukrit Kumar enhanced the flashinfer-ai/flashinfer repository by developing reproducible and CUDA graph-ready sampling functions for deep learning inference. He introduced optional seed and offset parameters to the sampling APIs, enabling deterministic random number generation and facilitating CUDA graph replay for efficient GPU execution. Using Python, CUDA, and PyTorch, Sukrit implemented comprehensive tests to verify both reproducibility with fixed seeds and variability with different values, ensuring robust behavior across scenarios. His work aligned with repository quality standards, passing pre-commit checks and updating tests, and laid a foundation for stable, graph-friendly inference pipelines in production machine learning workflows.